This is the code for reproducing our solution for the 2nd CVPR DataCV Challenge.
We have tested the following versions of OS and software:
-
Nvidia drivers: 470.57.02
-
1 x NVIDIA Tesla V100 SXM2 32GB
-
Python 3.8.19
-
pytorch 1.12.1
-
mmdet: 2.26.0
-
mmcv-full 1.7.0
- To create an environment to reproduce our best results, after cloning this work, first set up the conda environment based on instruction of the baseline project, but note that Python version here should be 3.8.
conda create -n tss python=3.8 -y
conda activate tss- Installing PyTorch 1.12 is required. Please adjust the CUDA version to match the one installed on your system.
pip install torch==1.12.0+cu116 torchvision==0.13.0+cu116 torchaudio==0.12.0 --extra-index-url https://download.pytorch.org/whl/cu116- Setting up the mmcv and mmdetection library .
pip install -U openmim
mim install mmcv-full==1.7.0
pip install yapf==0.40.1
cd mmdetection/
pip install -v -e . - And we also use YOLOv8, therefore you also need to install ultralytics,
pip install terminaltables pycocotools ultralytics==8.1.36 pillow==9.5.0-
Download region_100.zip and source_pool.zip via links from
https://github.com/yorkeyao/DataCV2024/tree/main -
Download the standard coco annotation of train2017: "instances_train2017.json"
wget http://images.cocodataset.org/annotations/annotations_trainval2017.zip -O /data/vdu2024/source_pool/annotations_trainval2017.zip - Extract zip file into "/data/vdu2024/", you may need to create this directory if not exists.
cd /data/vdu2024/source_pool
unzip annotations_trainval2017.zip- Move images from "/data/vdu2024/region100/train/001-100/***.jpg" into "train/" for YOLOv8 to predict.
cd /data/vdu2024/region100/train
mv */*.jpg ./- Update the annotation of coco for this competition. The coco_update.py is in the project's main directory. Then it will generate the "/data/vdu2024/source_pool/coco_updated.json" for the competition.
python coco_update.py- Download YOLOv8x checkpoint and infer on train set of Region 100. After inference, it will generate peusdo labels under "./runs/predict/labels" from the main directory. Move "predict" folder for label format conversion.
wget https://github.com/ultralytics/assets/releases/download/v8.1.0/yolov8x.pt
yolo detect predict model=./yolov8x.pt source=/data/vdu2024/region_100/train conf=0.1 imgsz=1280 save_txt=True classes=[2,5,7] save=False
mv runs/detect/predict ap_sort/- Run "yolo2coco.py" to convert YOLO-formatted labels into COCO-formatted json file, "train_yolov8.json" under "/data/vdu2024/region_100/".
python ap_sort/yolo2coco.py2. Training RetinaNet with Peusdo Labels with train set of Region 100 and inferring on COCO dataset.
python tools/train.py configs_tss/retinanet/001_train2source.py
python tools/test.py configs_tss/retinanet/001_train2source.py work_dirs/001_train2source/latest.pth --format-only --options "jsonfile_prefix=./"
mv .bbox.json ap_sort/infer_coco.jsonpython ap_sort/ap_sort_coco.py python tools/train.py configs_tss/retinanet/002_coco_8000.pypython mosaic.py
python tools/train.py configs_tss/retinanet/003_coco_8000_mosaic.pyDue to a severe crash of our server, we have unfortunately lost the original model weights referenced in the paper. As a result, the performance metrics associated with the "checkpoints.zip" from release differ slightly from those previously reported.
| Models | testA | testB |
|---|---|---|
| AP-COCO (paper) | 22.96 | 22.57 |
| AP-COCO (This repo) | 23.2 | N/A |
| AP-COCO+Mosaic (paper) | 22.62 | 22.85 |
| AP-COCO+Mosaic (This repo) | 23.1 | N/A |
To skip dataset selection process, you could directly load coco_8000.json by downloading "coco_8000.zip" from releases, extract it into "/data/vdu2024/source_pool/" and launch training code.
wget https://github.com/welovecv/datacv/releases/download/v1.0.0/coco_8000.zip -O /data/vdu2024/source_pool/coco_8000.zip
cd /data/vdu2024/source_pool/
unzip coco_8000.zip
python tools/train.py configs_tss/retinanet/002_coco_8000.pyOur code implementation is mainly built on baseline code given by baseline project. We appreciate 2nd DataCV Challenge competition organizers for their organization. We would also like to appreciate YOLOv8 for their works.